v0.11.0 Highlights
Product updates
Chat, voice, and stage
- You can now stop AIRI's speech manually from the stage UI, including the desktop and mobile layouts.
- Tool calls can now be retried from chat, so failed extension/gamelet actions do not always require starting over.
- Chat sync is more reliable when moving between desktop stage routes.
- Speech settings now support Voice Pack model selection, recommended voices, and better handling for "no speech" setups.
- Voice input and transcription flows were cleaned up, including hearing controls, audio device handling, and streaming transcription behavior.
- Chat actions now give clearer copy feedback, and desktop connection settings now expose copy-friendly connection details.
New providers and models
- Google Gemini TTS is now available as a speech provider.
- AIRI's official speech stack now includes Aliyun NLS ASR, StepFun TTS, default TTS model handling, generated voice previews, and official provider catalog routing.
- Vision provider settings were added for providers such as Ollama, LM Studio, Cloudflare Workers AI, Azure AI Foundry, Amazon Bedrock, and official providers.
Avatars and stage rendering
- Spine 2D model support is now integrated, with model settings, preview support, animation handling, and runtime validation.
- Live2D and VRM behavior is smoother: lip sync is less likely to be overridden by motions, Live2D zip loading is more reliable, and several missing default values were restored.
- The experimental Godot stage sidecar gained VRM runtime import, camera control, a new visual baseline, avatar glow, and rim lighting.
Desktop fixes
- Desktop now prevents duplicate app instances.
- Language choice persists correctly in the Electron app.
- Tray resizing now anchors to the dominant display.
- The updater path and exit behavior were fixed.
- Several widget, Spine preview, tooltip, Unicode, and hearing-control issues were cleaned up.
Account and billing
- Account linking now works better across different email providers, with clearer OAuth account selection and error handling.
- The account UI moved to accounts.airi.build.
- FLUX checkout from the Electron desktop renderer was unblocked.
- Billing for server-side usage is stricter and more reliable, including partial-debit handling to prevent unpaid usage gaps.
Fixes worth scanning for
- Provider validation now only applies to providers that are actually listed.
- Provider source filters now use corrected categories.
- Artistry now defaults to no provider instead of selecting one unexpectedly.
- Character card module selections now sync back to the active character card.
- Streaming TTS requests are gated more safely, and TTS session isolation was hardened.
- Official voice preview URLs are now exposed correctly.
- Auth callback trust handling was tightened.
- Social-only email sign-in hints are clearer.
To developers
- The server now has an in-process LLM/TTS gateway with official alias routing, upstream attempt propagation, structured OpenAI-compatible request logging, and streaming TTS catalog support.
- Server health checks moved to Kubernetes-style /livez and /readyz probes.
- The official provider catalog is now managed server-side, including catalog gating and official ASR/LLM alias policy.
- better-ws was added and server runtime WebSocket handling was refactored around liveness, routing, codecs, and response handling.
- Plugin terminology and APIs moved toward "extension" naming, with new extension entrypoint/session APIs and updated SDK surfaces.
- Gamelet/widget integrations now support coordinated iframe requests and retriable tool calls.
- The DashScope CosyVoice adapter moved to the v2 two-step REST flow.
- Docker builds now install needed build dependencies and run package build steps more reliably.
To contributors
- cap-vite now auto-selects the first available native device when no target is specified.
- Playwright was upgraded to avoid install hangs on Node 24.
- CI now repairs pnpm lockfile snapshots, automates sponsor-update PRs, mirrors release assets to GitCode, and includes Godot stage publishing.
- The workspace dependency setup moved further into catalogs.
- Docs were refreshed with desktop quick-start updates, install tips, AtomGit acknowledgements, and a character card template reference.
- Computer-use MCP gained terminal screen heuristics, bounded terminal capture, background scheduler improvements, browser repair suggestions, and safer cleanup for timed-out terminal processes.
- The Minecraft service gained stronger perception, reflex, combat, and planning behavior.
Upgrade notes
- Server operators should update health checks from legacy endpoints to /livez and /readyz.
- Deployments using the old auth UI host should move to accounts.airi.build.
- Server TTS deployments should review the newer UNSPEECH_UPSTREAM configuration and official provider catalog setup.
- Extension/plugin authors should review the plugin-to-extension rename and update manifests/APIs where applicable.
